cross-validation ANN

Hii...

I want to ask how to divide our data for training, validation, and testing for ANN in MATLAB, specifically for cross-validation.

Thank you

Answer 1

[Greg Heath]

What is your goal?...

Find the single best network? If so, what is your criterion?

Find the M best networks to use in a parallel configuration?

The emphasis on the Ntrn/Nval/Ntst split ratio is misleading. The most important things are the actual magnitudes.

Ntrn should be large enough to obtain sufficiently accurate weight estimates.

Nval should be large enough determine a satisfactory set of training parameters (No. of hidden nodes, epochs, etc).

Ntst should be large enough to predict a sufficiently precise estimate of errors on unseen operational data.

There are rules of thumb for independently choosing these values. However, the easiest approach is to just use trial and error by first making multiple (e.g., 10) runs with the default value. If unsatisfactory, Increase Ntrn with Ntst = Nval.

Hope this helps.

Greg

[Muammar]

Hii..thank you, but I just start learning MATLAB, therefore I cant understand your answer fully.. but I will try. Once again, Thank you very much for your help.